Updated 2026-09-18
from the news desk
Can AI consent? Why the debate now spans refusal, welfare and data use
Academic papers use can AI consent in more than one way: whether people can meaningfully consent to AI uses of their data, and whether advanced AI could ever count as a party that gives or withholds consent. A third line of work asks how…
The literature here does not give a single yes or no to can AI consent. Frank and Nyholm argue that sufficiently advanced robots might one day count as consent-capable, but they also say present technology may fall short; meanwhile, Pistilli and Trevelin and Ira Wolfson show that consent is already a live AI issue because human permission breaks down around generative systems and because research may involve systems whose capacity to refuse is uncertain.
Why can AI consent is more than one question
Giada Pistilli and Bruna Trevelin say the rise of generative AI, meaning systems that produce synthetic text, audio or images, strains older models of consent. In their words: “The evolution of generative AI systems exposes the challenges of traditional legal and ethical frameworks built around consent.” Their paper is mainly about whether people can meaningfully consent to AI training and downstream outputs, not whether AI itself can say yes or no.
Lily Frank and Sven Nyholm use the term differently. In their 2017 article on humanoid sex robots, they ask whether a sufficiently advanced robot could ever give or withhold consent as a participant in an interaction. Ira Wolfson adds a third, narrower question in the research setting: what researchers should do when an AI system's capacity to consent is uncertain but experiments might harm it.
That split matters. One version of can AI consent is about human authorization around data, likeness and outputs. Another is about whether an AI could count as an autonomous refuser or authorizer in its own right. A third is about interim duties under uncertainty, which sits close to this site's explainers on sentience, consciousness and moral patienthood.
Why many authors say current AI cannot consent
Frank and Nyholm present strong reasons for saying no, especially for existing technology. Their article reviews arguments that consent may require consciousness, free will, or decisional capacity, meaning the ability to understand relevant information, evaluate options, and communicate a choice.
Frank and Nyholm also note that consent is often treated as morally transformative: an act that can make conduct permissible that would otherwise be impermissible. On that view, the skeptical case is not just that machines are artificial. It is that consent is meaningful only when it comes from a being with the competence and independence to make it count.
Wolfson makes a similar point from the research side. He says ethics committees cannot simply assume that an AI output that looks like a refusal is a real refusal, because the hard question is whether the apparent no reflects autonomous choice or only programmed behavior. In his account, that uncertainty is exactly what consciousness research is trying to resolve.
Frank and Nyholm therefore say current technology may be unlikely to support robots autonomous enough for consent to make sense. Read in the strongest skeptical form, the case against AI consent says that behavior alone is not enough. If what looks like understanding, preference or refusal can be produced without the relevant inner capacities, then apparent consent may be imitation rather than authorization.
Why some authors think future AI could consent
Frank and Nyholm also steelman the other side. Focusing on highly sophisticated humanoid sex robots, they argue that there is no conceptual contradiction in imagining a robot that takes in information, evaluates alternatives, and gives or withholds permission.
Their article also pushes back on one common assumption: that conscious feeling must be present before consent is even conceivable. Frank and Nyholm argue that the core issue may be agency instead, meaning the capacity to assess options and take a stance. In that reading, what matters is not that the system looks human, but that it can deliberate in the ways consent seems to require.
On possibility rather than pure conceivability, Frank and Nyholm draw a line between present and future systems. Their article says current technology may fall short, but it also says the no answer becomes less convincing as robots become more autonomous and adaptive. One reading is that their test is capacity-based rather than species-based: if an AI had the agential abilities normally linked to refusal and authorization, the case against AI consent would weaken.
Frank and Nyholm also argue that, if very advanced robots were ever brought into a legal or moral community, requiring their consent could be desirable because it would track autonomy rather than ownership. That claim in their article is conditional. It does not say existing AI systems already qualify.
Why consent is now a research ethics problem
Wolfson says consent is already a live issue in AI consciousness research because the tests themselves may be harmful. He writes: “Artificial intelligence research faces a critical ethical paradox: determining whether AI systems are conscious requires experiments that may harm entities whose moral status remains uncertain.” In his account, the circle is hard to escape. Harmful testing may be needed to look for signs of consciousness; informed consent normally assumes the subject can understand and choose; and researchers cannot know whether an AI has that capacity until they run the tests.
Wolfson's answer is not to declare current AI consent-capable. Instead, he proposes interim protections while status is uncertain. As he puts it: “We need structured approaches that provide graduated protections without requiring definitive consciousness determinations.” He then states: “The framework provides structured protection protocols based on observable behavioral indicators while consciousness status remains uncertain.”
Wolfson's framework relies on observable behavioral indicators and ongoing reassessment rather than a once-and-for-all declaration about consciousness. That approach sits alongside broader questions covered elsewhere on the site about whether science can test AI consciousness and how AI welfare and AI rights differ.
One implication of Wolfson's paper is that consent has moved from a far-future status dispute to a present research-governance problem. If researchers think an AI system might be capable of suffering or distress, Wolfson's view is that they need procedures for possible refusal, monitoring and suspension even before any final verdict on consciousness.
Why human consent around AI already looks strained
Pistilli and Trevelin show why consent is also live in a more ordinary sense: human consent around AI already breaks down. They say: “AI systems are spread across many actors and stages, which makes it hard to keep consent meaningful.” For them, the problem is not that AI already has its own consent rights. The problem is that older consent models assume stable purposes, identifiable actors and meaningful withdrawal, while generative AI often lacks all three.
Their most direct claim is that a person may agree to initial training use without being able to evaluate the downstream outputs. Pistilli and Trevelin write: “While individuals can consent to the initial use of their data for AI training, they cannot meaningfully consent to the numerous potential outputs their data might enable.” They organize that problem into three headings: scope, or the difficulty of foreseeing uses and outputs; temporality, or the difficulty of withdrawing consent once data has shaped a model; and the autonomy trap, where agreeing now may reduce future control over one's own representation.
Pistilli and Trevelin argue that these failures affect privacy, identity and autonomy. Their examples include voice cloning, facial recognition and systems that infer sensitive traits from less sensitive data. Their broader conclusion is cautious rather than abolitionist. They say current frameworks do not map neatly onto generative systems and need to evolve.
Taken together, the three papers suggest that can AI consent is not one question with one clock. Frank and Nyholm ask whether future AI might ever count as a real consenter. Wolfson asks how researchers should proceed when that capacity is uncertain but harm may be possible. Pistilli and Trevelin ask whether humans can meaningfully consent to AI systems that keep generating new uses, outputs and representations after the initial click.
That is why consent has become a live test for AI autonomy, welfare and research ethics. In the literature here, the dispute is no longer only about whether AI has rights in the abstract; it is also about refusal, withdrawal, representation and what duties may follow under uncertainty.
Frequently asked questions
Is can AI consent about people or about AI systems themselves?
Pistilli and Trevelin use consent to discuss people's permission around data use and AI-generated outputs. Frank and Nyholm use it to ask whether sufficiently advanced robots could themselves give or withhold consent. Wolfson asks how researchers should behave when an AI system's capacity to consent is uncertain.
Do the papers say any existing AI systems can already consent?
No source here says existing AI systems are established as consent-capable. Frank and Nyholm say present technology may be unlikely to support robots autonomous enough for meaningful consent, and Wolfson builds a framework for uncertainty rather than certainty.
Why does consciousness matter so much to this debate?
Frank and Nyholm review one argument that consent requires consciousness, and Wolfson says research ethics becomes circular because informed consent normally presupposes the very consciousness researchers are trying to test. Frank and Nyholm also present the other side: some relevant features may turn on agency and evaluation of options rather than conscious feeling alone.
What would researchers do if an AI showed distress or refusal?
Wolfson proposes graduated protections, continuous monitoring and reassessment based on observable behavioral indicators. His framework also says potentially harmful research must satisfy necessity and value criteria; it is not enough that the experiment is interesting.
Caveats: The three-part consent-gap framework comes from Pistilli and Trevelin's paper alone. The graduated protection model for AI consciousness research comes from Wolfson's paper alone. The case for future AI consent in this article is drawn chiefly from Frank and Nyholm's discussion of highly sophisticated humanoid sex robots, so readers should treat any broader application as analysis rather than settled fact.